RBF Collocation - · PDF fileRBF collocation C.T.Mouat ∗andR.K.Beatson...
Transcript of RBF Collocation - · PDF fileRBF collocation C.T.Mouat ∗andR.K.Beatson...
RBF Collocation
C. T. Mouat and R. K. Beatson
Department of Mathematics & Statistics,
University of Canterbury,
Private Bag 4800, Christchurch, New Zealand.
Report Number: UCDMS2002/3 February 2002
Keywords: radial basis functions, domain decomposition, collocation, pde
RBF collocation
C. T. Mouat∗ and R. K. Beatson∗
version of February 28, 2002
1 Introduction
In recent years radial basis function collocation has become a useful alternative to finite
difference and finite element methods for solving elliptic partial differential equations. RBF
collocation methods have been shown numerically (see for example [17]) and theoretically
(see [14, 13]) to be very accurate even for a small number of collocation points. In appli-
cation finite difference methods often have a low approximation order and consequently
can require a large grid and considerable computation to obtain a sufficiently accurate
solution. RBF collocation has been applied to linear elliptic PDEs in R2 and R3 [18], to
time dependent problems [15, 16], and to non-linear problems [10].
In this paper we present new numerical results for RBF collocation. These results
show that collocation with a basic function from the Matern class can be more accurate
than collocation with the multiquadric basic function. Also, we present and implement an
algorithm which solves linear and non-linear collocation equations with the multiquadric
when N is large and c < 2/√N .
Section 2 briefly outlines RBF collocation and discusses difficulties with the method.
These include poor conditioning and full matrices when using globally supported basic
functions, and lower accuracy when using compactly supported basic functions. In Section
1Department of Mathematics and Statistics, University of Canterbury, Christchurch, New Zealand.
2
c. t. mouat and r. k. beatson 3
3 we give numerical results using a family of basic functions known as the Matern family.
These numerical results show that this family is an effective alternative to the multiquadric
basic function in many situations. Finally, in the last sections, we present a method which
can be used to solve large collocation problems with the multiquadric basic function and
c < 2/√N . Numerical experiments on linear PDEs show convergence to the solution for
small enough values of c. It is hoped that in the future the algorithm will be able to be
applied to larger values of c. This new algorithm combines the use of approximate cardinal
functions and domain decomposition to iteratively find the solution of the collocation
problem. Using approximate cardinal functions as a change of basis has been shown to be
effective in the interpolation setting [2]. Previously solving a collocation system required
O(N3) operations (for globally supported Φ) and was not possible for large N . The method
presented here solves the collocation system in O(N logN) operations if c < 2/√N and
the PDE is suitable.
2 RBF collocation
This paper considers solving a suitable elliptic PDE of the form
Lu = f in Ω ⊂ Rd, (1)
u = g in ∂Ω,
by radial basis function collocation. In (1) f, g : Rd → R are known and ∂Ω is the
boundary of the region Ω. L is a differential operator and may be linear or non-linear. If
L is non-linear a multilevel Newton iteration is required and a linearized system is solved
at each level.
The unknown solution, u, to the PDE is approximated by a radial basis function, uφ,
of the form
uφ(·) = p(·) +N∑
j=1
λjΦ(· − xj). (2)
4 rbf collocation
Here λ = [λ1, . . . , λN ]T are coefficients to be found, p ∈ πd
k, and Φ is a basic function, such
as the multiquadric. If L is time dependent then we let λ be a function of time and solve
for λ(t) at a finite number of discrete time steps. For more discussion on this case see [17].
For the moment assume L is time independent. Now for uφ to satisfy the PDE (1) then
Luφ(x) = f(x), x ∈ Ω,uφ(x) = g(x), x ∈ ∂Ω. (3)
Clearly this cannot generally be achieved for every point in Ω. By choosing N distinct
collocation points XI = x1, . . . , xNI ⊂ Ω and XB = xNI+1, . . . , xN ⊂ ∂Ω and ensuring
(3) holds at these points we expect uφ will be a good approximation to u. For the choice
of uφ in (2) the collocation equations are
Lp(xi) +N∑
j=1
λjLΦ(xi − xj) = fi, i = 1, . . . , NI ,
p(xi) +N∑
j=1
λjΦ(xi − xj) = gi, i = NI + 1, . . . , N,
along with the side conditions
N∑j=1
λjq(xj) = 0, for all q ∈ πdk.
This leads to the equivalent matrix form
WL PL
WB PB
P T O
λ
a
=
f
g
0
, (4)
where
(WL)ij = LΦ(xi − xj), xi ∈ XI , xj ∈ X,
(WB)i−NI ,j = Φ(xi − xj), xi ∈ XB, xj ∈ X,
(PL)ij = Lpj(xi), xi ∈ XI ,
(PB)i−NI ,j = pj(xi), xi ∈ XB,
(5)
c. t. mouat and r. k. beatson 5
and p1, . . . , pdim(πdk) forms a basis for πd
k. The vector a consists of coefficients with respect
to this basis. Solving this collocation system for the coefficients [λT aT ]T , when N is large,
is the emphasis of later sections of this paper. The strategy there is to precondition the
collocation matrix
A =
WL PL
WB PB
P T O
, (6)
so that the preconditioned system is solved quickly using an iterative method. The collo-
cation matrix, A, in (6) has not been proven to be non-singular but in [20] it was shown
that finding a numerically singular matrix was very rare. The positioning of the centres
has an effect on the accuracy of RBF collocation. However, to keep the discussion simpler,
we only consider gridded centres.
Equation (2) is the form of the RBF approximation that was initially presented by
Kansa [17]. This form is often called unsymmetric collocation due to the matrix in (6)
being unsymmetric. An alternative approach [8], referred to as symmetric collocation,
takes the form
uφ(·) = p(·) +NI∑j=1
λjLΦ(· − xj) +N∑
j=NI+1
λjΦ(· − xj), (7)
where L is the operator L now applied to the second argument, xj. Note that the absolute
values of LΦ(y − x) and LΦ(y − x) are equal for any x and y. For the choice of uφ in (7)
the collocation equations lead to the interpolation system
WLL WL PL
W TL
WB PB
P TL P T
B O
λ
a
=
f
g
0
. (8)
6 rbf collocation
The matrices in (8) are,
(WLL)ij = LLΦ(xi − xj), xi, xj ∈ XI ,
(WL)i,j−NI= LΦ(xi − xj), xi ∈ XI , xj ∈ XB,
(WL)i,j−NI= LΦ(xi − xj), xi ∈ XI , xj ∈ XB,
(WB)i−NI ,j−NI= Φ(xi − xj), xi, xj ∈ XB,
(9)
and PL and PB are the same as in (4). The main advantage of this formulation is that
it is provably non-singular (see [8, 22]). However the RBF in (7) is not as widely used as
Kansa’s original due to an extra application of L requiring that Φ be more differentiable.
For nonlinear collocation using (7) also increases the complexity of the method. Some
numerical results comparing the two approaches can be found in [8].
Both collocation systems are generally very badly conditioned which can restrict the use
of RBF collocation to systems with only a few thousand centres. Theoretical results show
that multiquadric interpolation becomes more accurate as the multiquadric parameter c
increases [19]. A lot of numerical evidence agrees with this in the collocation setting.
However, as c gets larger the graph of the basic function becomes flatter and this leads to
bad conditioning. Thus as the accuracy of the approximation increases then often so does
the ill-conditioning. Various techniques have been used with mixed success to combat this
problem (see for example [18]).
The problems associated with using globally supported basic functions have led to the
use of compactly supported basic functions such as the Wendland functions [21]. If the
support is small then matrix-vector multiplies can be calculated in O(N) operations. Theproblem with compactly supported basic functions is that good approximations to the
solution are only obtained when the support is large. For accurate results the sparcity of
the matrix is lost. A multilevel approach with smoothing can improve the accuracy of the
RBF approximation [7] but multiquadric basic functions are usually more accurate.
c. t. mouat and r. k. beatson 7
3 Collocation with Matern basic functions
Traditionally multiquadric or compactly supported basic functions are the preferred choice
for RBF collocation. Numerical evidence has shown good results with these choices of
basic functions for various types of problems. Other alternatives that are common in the
RBF interpolation setting can be restricted in their use for collocation. For example, the
Laplacian of the thin-plate spline is
∆Φ(x) = 4 log(‖x‖) + 4,
which has a discontinuity at zero. The Laplacian of the exponential basic function also
has a discontinuity at zero. This makes the use of the thin-plate spline and exponential
limited in RBF collocation.
Due to the conditioning problems associated with the multiquadric we consider the use
of alternative basic functions for RBF collocation. This section presents numerical results
for some simple PDEs using the Matern family as basic functions. The Matern family is
given by
φν(r) =21−ν
Γ(ν)(cr)νKν(cr), (10)
where Kν is a modified Bessel function of order ν > 0 (note that ν is also a smoothness
parameter) and c > 0. If n is a nonnegative integer then (10) simplifies to
φn+1/2(r) =exp(−cr)(cr)n
(2n− 1)!!n∑
k=0
(n+ k)!
k!(n− k)!(2cr)k.
Some examples for various values of ν are:
ν = 1/2, φ(r) = exp(−cr),
ν = 1, φ(r) = crK1(cr),
ν = 3/2, φ(r) = (1 + cr) exp(−cr),
ν = 5/2, φ(r) = (1 + cr + c2r2/3) exp(−cr).
(11)
Although we only consider unsymmetric collocation here the motivation behind the use
of the Matern class comes from the results of Franke and Schaback [14, 13] in the symmetric
8 rbf collocation
collocation setting. They show that for a PDE of order m the L∞ approximation order
for RBF collocation with a Matern basic function will be ν − m. Note that this result
is for solutions u in the “native space” of Φ. A complete review of the work of Franke
and Schaback is beyond the scope of this thesis but the reader is referred to their papers
[14, 13].
Tables 1 and 2 contain condition numbers of the collocation matrix (6) and relative
error results for the PDE
∆u = 32 cos(4x1 + 4x2), (x1, x2) ∈ Ω,u = cos(4x1 + 4x2), (x1, x2) ∈ ∂Ω,
where Ω is the unit square. The relative error is ‖s − u‖2/‖u‖2 where s is the values
of the RBF and u the values of the true solution evaluated on a uniform grid of size
(2√N − 1) × (2
√N − 1). The basic functions we compare are the multiquadric and the
Matern, ν = 9/2, function.
It is clear from the tables that as the basic function becomes flatter the condition
number increases for a fixed set size. In the case of the multiquadric this corresponds to c
increasing, whereas for the Matern function this corresponds to a decrease in c.
Table 1 shows results for centres on a uniform grid in [0, 1]2. The smallest relative
error for the Matern function is about 9 times smaller then the smallest relative error for
the multiquadric. However, both these experiments have condition numbers greater than
1020. If we look at experiments with condition numbers that are about 1016 or less then
the difference between the basic functions is even more dramatic. The best results are
then approximately 1.7 × 10−5 and 4 × 10−7 for the multiquadric and Matern functions
respectively. The error for the Matern function is about 40 times smaller than the error
for the multiquadric!
The same experiments were repeated on a grid of shifted Chebychev nodes in [0, 1]2.
The results are in Table 2. The errors for these trials were as low as 1.13×10−8 for Matern
collocation on 4225 centres. Overall, for this PDE, RBF collocation with the Matern basic
function was more accurate than RBF collocation with the multiquadric especially for
c. t. mouat and r. k. beatson 9
large N . Also collocation on the Chebychev grid was more accurate than collocation on
the uniform grid.
Number Multiquadric Matern, ν = 9/2
of centres c relative condition c relative condition
N error number error number
15/9 7.552(-5) 6.555(17) 0.5 1.056(-3) 1.694(16)
13/9 1.217(-4) 1.038(17) 1.0 1.079(-3) 3.069(13)
9× 9 11/9 2.582(-4) 2.037(15) 1.5 1.108(-3) 9.406(11)
9/9 5.684(-4) 2.877(13) 2.0 1.163(-3) 9.117(10)
7/9 1.250(-3) 2.447(11) 2.5 1.251(-3) 1.601(10)
15/17 2.095(-6) 1.492(19) 0.5 6.860(-4) 6.704(18)
13/17 5.656(-6) 3.610(19) 1.0 4.488(-5) 1.603(17)
17× 17 11/17 1.854(-5) 5.778(18) 1.5 4.703(-5) 3.023(15)
9/17 6.131(-5) 2.378(16) 2.0 5.175(-5) 3.109(14)
7/17 2.152(-4) 4.151(13) 2.5 5.796(-5) 5.518(13)
15/33 1.895(-6) 2.594(20) 2.0 1.628(-6) 1.338(19)
13/33 9.995(-7) 4.394(20) 2.5 1.952(-6) 1.484(17)
33× 33 11/33 3.345(-6) 9.664(20) 3.0 2.266(-6) 3.162(16)
9/33 1.397(-5) 2.063(18) 3.5 2.684(-6) 9.402(15)
7/33 5.648(-5) 1.397(15) 4.0 3.232(-6) 3.188(15)
15/65 1.536(-6) 1.862(21) 3.0 2.139(-7) 3.682(20)
13/65 1.099(-6) 3.087(21) 4.0 1.112(-7) 1.472(20)
65× 65 11/65 1.423(-6) 8.629(20) 5.0 1.419(-7) 3.792(18)
9/65 4.170(-6) 7.398(20) 6.0 2.290(-7) 8.410(17)
7/65 1.697(-5) 2.086(16) 7.0 3.724(-7) 5.944(16)
Table 1: Radial basis function collocation of the Poisson equation with solution cos(4x1 +
4x2). Uniform grid.
10 rbf collocation
Number Multiquadric Matern, ν = 9/2
of centres c relative condition c relative condition
N error number error number
15/9 3.026(-5) 4.737(17) 0.5 3.966(-4) 1.775(16)
13/9 4.962(-5) 2.614(16) 1.0 4.164(-4) 3.464(13)
9× 9 11/9 1.038(-4) 6.360(14) 1.5 4.342(-4) 1.085(12)
9/9 2.307(-4) 1.102(13) 2.0 4.634(-4) 1.059(11)
7/9 5.189(-4) 1.305(11) 2.5 5.066(-4) 1.864(10)
15/17 4.469(-5) 1.454(19) 1.0 4.136(-6) 4.806(18)
13/17 4.756(-7) 7.760(18) 2.0 2.498(-6) 4.190(15)
17× 17 11/17 7.822(-7) 5.132(18) 3.0 3.370(-6) 1.732(14)
9/17 2.924(-6) 9.153(16) 4.0 4.865(-6) 1.725(13)
7/17 1.344(-5) 5.349(14) 5.0 7.407(-6) 2.721(12)
11/33 6.382(-6) 9.656(19) 4.0 1.222(-7) 1.957(20)
9/33 2.256(-5) 6.774(19) 5.0 1.554(-8) 2.005(18)
33× 33 7/33 1.082(-6) 3.598(19) 6.0 2.389(-8) 1.843(17)
5/33 4.329(-6) 6.523(17) 7.0 3.878(-8) 3.591(16)
3/33 1.549(-5) 7.397(13) 8.0 6.210(-8) 1.226(16)
5/65 9.659(-7) 1.204(22) 15 1.213(-8) 6.011(20)
4/65 7.014(-7) 4.365(20) 20 1.127(-8) 3.828(18)
65× 65 3/65 2.533(-6) 1.893(20) 25 4.414(-8) 5.882(17)
2/65 1.105(-4) 4.786(16) 30 2.111(-7) 1.131(17)
1/65 4.847(-3) 3.162(12) 35 9.722(-7) 3.879(16)
Table 2: Radial basis function collocation of the Poisson equation with solution cos(4x1 +
4x2). Chebychev grid.
c. t. mouat and r. k. beatson 11
4 Solving the collocation system for large N
For globally supported basic functions directly solving the collocation system (4) requires
O(N3) operations and O(N2) storage without using any customised method. This sec-
tion presents a new algorithm for solving this system in O(N logN) operations and O(N)storage with the multiquadric basic function. The algorithm uses a change of basis pre-
conditioner in conjunction with domain decomposition and a fast matrix-vector multiply.
The greatest computational cost at each iteration is at least one matrix-vector multiply.
Fast matrix-vector product algorithms allow this to be achieved in O(N logN) operations
using a suitable fast evaluation code. These fast evaluation codes exist for a variety of
functions [1, 3, 5].
4.1 Domain decomposition
This subsection considers a domain decomposition algorithm for centres separated into
two domains. Without loss of generality refer to XI = x1, . . . , xNI as good points and
XB = xNI+1, . . . , xN as bad points. Also assume that NI >> |XB| =: NB. Then if an
interpolant, s, is of the form
s(·) =N∑
j=1
λjΨ(· − xj),
where the coefficients λj are to be found then the interpolation matrix is Bij = Ψ(xi −xj).
This can be split into the form
B =
BII BIB
BBI BBB
. (12)
In (12) Bjk, j, k ∈ I, B has size Nj × Nk and is the matrix from evaluating the Ψs
centred at Xk, at points in Xj. Now applying a simple domain decomposition algorithm to
this system we can iteratively obtain a solution. This method is given by Algorithm 4.1.
The notation rB, rI refers to the residuals restricted to centres XB and XI respectively.
12 rbf collocation
Algorithm 4.1 domain− decomposition(X, f,NI)
SETUP
1. Create a set, XI , of good points and a set, XB of bad points
2. Form Ψ elements for each point x in X
3. r ← f and s ← 0
ITERATIVE SOLUTION
1. while ‖r‖ > ε
2. Solve for the coefficients µ of a bad point approximation via direct or
approximate solutions of BBBµ = rB
3. sbad ← ∑j:xj∈XB
µjΨj
4. Evaluate rI = rI − sbad(XI)
5. Solve for the coefficients µ of a good point approximation via approx-
imate solution of BIIµ = rI
6. sgood ← ∑j:xj∈XI
µjΨj
7. Update the RBF s = s+ sbad + sgood
8. Update the residual r = f − s(X)
9. end while
At the beginning of each iteration coefficients µ are found so that
∑j:xj∈XB
µjΨj(xi) = ri, xi ∈ XB.
c. t. mouat and r. k. beatson 13
Because NB is small compared to NI this is relatively efficient to solve. The residuals
are then updated and a similar system on the good points is solved. Although NI is
large we can solve the system on the good points efficiently by GMRES if the eigenvalues
of BII are sufficiently clustered. Each GMRES iteration will require the computational
cost of a matrix-vector multiply. In our experience an exact solution at step 5 is not
required. Instead reducing the residual by a few orders of magnitude will suffice. If BII
is an approximation to the identity then most off diagonal elements will be near zero.
An approximation to BII can easily be found by retaining only a small number, say σ,
of the largest magnitude entries per column. A matrix-vector product will then only
require O(σN) operations instead of the O(N logN) required using a fast matrix-vector
code. Numerical evidence shows that this approximation increases the number of outer
iterations by less then four times but significantly decreases the total number ofO(N logN)
matrix-vector products (which is the main computational cost of the algorithm).
The final step is to update the residual which can be achieved in O(N logN) operations.
Note that all matrix-vector multiplies will be O(N logN) only if a suitable fast evaluation
algorithm exists for the basic functions Φ and LΦ. For example, if L is the Laplacian and
Φ the multiquadric then
LΦ(·) = ‖ · ‖2 + 2c2
(‖ · ‖2 + c2)3/2=
1
(‖ · ‖2 + c2)1/2+
c2
(‖ · ‖2 + c2)3/2,
which is a combination of two members of the multiquadric family. Fast evaluators are
available for functions of this type [3].
Algorithm 4.1 should be modified to include a coarse grid correction at each iteration.
We usually take the number of points in the coarse grid to be about NB.
4.2 Approximate cardinal functions
In the previous section it was assumed that the matrix BII had clustered eigenvalues. In
this section we achieve this by forming Ψ elements as approximate cardinal functions. We
also explain why this approach doesn’t work for large values of the multiquadric parameter
c. Using approximate cardinal functions as a change of basis has been shown to be effective
14 rbf collocation
in the interpolation setting [2, 4]. The main difference in the collocation case is that there
are different operators on the interior and on the boundary. If our aim was for B to be a
good approximation to the identity, as in the interpolation case, then the Ψj’s would be of
the form,
Ψj(xi) ≈ 0, xi ∈ XB,
LΨj(xi) ≈ 0, xi ∈ XI ,
along with one of the constraints,
LΨj(xj) = 1, if xj ∈ XI ,
Ψj(xj) = 1, if xj ∈ XB.
In our experience forming approximate cardinal functions to satisfy these conditions is
difficult. Instead we form approximate cardinal functions which ensure BII is a good
approximation to the identity and use the domain decomposition approach given in Algo-
rithm 4.1. The bad points are the boundary points and the good points are the interior
points. For a uniform distribution of points in R2 the number of boundary points, NB, is
proportional to N1/2 so direct solution of a linear system on these points requires O(N3/2)
operations. Calculating the LU factorisation to BBB as part of the setup means this cost
is only incurred once. Subsequent use of this LU decomposition to solve a system requires
O(N2B) = O(N) operations.Each Ψ element is of the form
Ψj(·) = pj(·) +∑i∈Sj
λjiΦ(· − xi),
where the set Sj is often a set of indices of the nearest β points to xj. For interior points
we construct Ψ elements so that
LΨj(xj) = 1,
LΨj(x) = O(‖x− xj‖−3) as ‖x− xj‖ → ∞.
c. t. mouat and r. k. beatson 15
Approximate cardinal functions of this type are referred to as decay element approximate
cardinal functions and are found by a constrained least squares problem as mentioned in
[2].
The set Xj is defined to be the centres in X such that xi ∈ Xj if and only if i ∈ Sj.
For boundary points use a pure local approach of the form
Ψj(xj) = 1
Ψj(xi) = 0, xi ∈ Xj ∩XB,
LΨj(xi) = 0, xi ∈ Xj ∩XI .
The pure local approximate cardinal functions are found by solving a collocation system
on |Sj| nodes for each j.
In our experience we have noticed that creating approximate cardinal functions is only
effective for c < 2/√N (if the centres are a uniform grid in [0, 1]2). In Figure 1 approxi-
mate cardinal functions formed using a decaying strategy for two different values of c are
compared. Clearly the Ψ element formed with the larger value of c is not a very good
approximation to a cardinal function. The required rate of decay is not achieved until
further away from the centre of the Ψ element. An explanation for this is the regions of
validity of the far field expansions. Consider finding a Ψ centred at xj and based on centres
with maximum distance H from xj. The Ψ element will decay in the region of validity of
the far field expansion of the cluster. This expansion is given in [3] and is valid outside the
circle ‖x − xj‖ =√H2 + c2. If c is large then the radius of this circle increases and the
region of validity is further from xj (see Figure 2 ) and thus the decay of Ψ occurs further
away.
16 rbf collocation
00.2
0.40.6
0.81
00.2
0.40.6
0.81
−0.2
0
0.2
0.4
0.6
0.8
1
(a) A Ψ function formed with c = 2/33.
00.2
0.40.6
0.81
00.2
0.40.6
0.81
−2
−1.5
−1
−0.5
0
0.5
1
1.5
(b) A Ψ function formed with c = 4/33.
Figure 1: Ψ elements based on fifty local centres for two different values of c. Notice that
the Ψ function decays quicker with the smaller value of c.
** *
*
*
**
0+ c 2 H 2 1
H
+ c 2 2 H 2
Figure 2: Far field expansion regions of validity for centres inside the circle ‖ · ‖ ≤ H and
multiquadric parameters c1, c2 with c2 > c1. The region of validity corresponding to c2 is
outside the dotted circle and for c1 is outside the dashed circle.
c. t. mouat and r. k. beatson 17
5 Numerical results
In this section we present numerical results for RBF collocation of linear PDEs of the
form (1). These results show good convergence of the algorithm when the multiquadric
parameter c is suitably small and constant. All numerical experiments are in the domain
[0, 1]2 with the collocation nodes forming an n× n grid. The iterations are stopped once
the relative 2-norm residual is less than 10−8.
To initially try this method we consider solving Poissons equation in R2 with the
solutions
f1(x1, x2) = exp (2x1 + 2x2) , (13)
f2(x1, x2) = exp(−1000((x1 − 1/2)6 + (x2 − 1/2)6)
). (14)
RBF collocation solutions for these two PDEs can be found in Figure 3. The results from
00.2
0.40.6
0.81
00.2
0.40.6
0.810
10
20
30
40
50
60
(a) Exact solution is f1.
00.2
0.40.6
0.81
00.2
0.40.6
0.81
0
0.2
0.4
0.6
0.8
1
(b) Exact solution is f2.
Figure 3: RBF collocation on a 33× 33 grid of centres in [0, 1]2 with c = 2/33.
applying Algorithm 4.1 to f1 and f2 are in Tables 3 and 4 respectively. The algorithm was
applied using both exact matrix-vector products and approximate matrix-vector products
at step 5. We will refer to these different implementations as Algorithm 4.1(a) and Al-
18 rbf collocation
gorithm 4.1(b) respectively. For both implementations exact matrix-vector products are
always used at step 7. The “matrix-vector” column in the tables give the total number
of exact matrix-vectors calculated to find the solution. For Algorithm 4.1(b) the total
number of exact matrix-vector products is equal to the number of outer iterations. The
“2-norm residual” column is the relative 2-norm residual ‖Aλ − f‖2/‖f‖2 where f is the
right hand side vector [fT gT ]T in (4). The tables show that the algorithm converges
for both small and large values of N . As expected, Algorithm 4.1(b) requires more outer
iterations to converge but the total number of exact matrix-vector products is reduced.
This is a sizeable computational saving for large N .
Overall the number of outer iterations remained fairly stationary for Algorithm 4.1(a).
When approximate matrix-vectors were used the number of outer iterations increased
slightly but not dramatically as N increased. Thus it would be feasible to solve even
larger systems using this algorithm.
From these experiments we can conclude that the algorithm will at least work on some
simple PDEs when c is small. Usually c is required to be large for higher accuracy but
in the case of f2 we noticed that c = 2/√N was nearly optimal for small data sets and
using a Matlab \ operator to solve the systems. Carlson and Foley [6] suggest that a smallshape parameter will be more accurate if the function values vary rapidly. The algorithm
presented here may therefore be more applicable for solutions of this type.
A suitably modified algorithm has shown promising results for the nonlinear PDE given
in [11, 10].
References
[1] R. K. Beatson and E. Chacko, Fast evaluation of radial basis functions: A multivariate
momentary evaluation scheme, to appear in Curve and Surface Fitting: Saint Malo
1999, A. Cohen, C. Rabut and L.L. Schumaker (eds), Vanderbilt Univ. Press, Nashville
(2000).
c. t. mouat and r. k. beatson 19
Exact matrix-vectors Approximate matrix-vectors
N NC Outer Matrix 2-norm Outer σ 2-norm
iterations vectors residual iterations residual
289 64 8 48 9.323(-9) 9 43.2 8.996(-9)
1089 121 8 48 1.413(-9) 14 44.6 6.745(-9)
4225 400 7 42 7.051(-9) 20 47.4 7.004(-9)
16641 625 9 54 2.409(-9) 27 48.2 8.539(-9)
Table 3: Results from Algorithm 4.1 on function f1. NC is the number of coarse grid points
and σ is the average number of non-zero elements per column in the approximation to B.
Exact matrix-vectors Approximate matrix-vectors
N NC Outer Matrix 2-norm Outer σ 2-norm
iterations vectors residual iterations residual
289 64 9 54 1.940(-9) 9 43.2 9.313(-9)
1089 121 7 42 1.576(-9) 16 44.6 8.487(-9)
4225 400 6 36 2.313(-9) 22 47.4 5.396(-9)
16641 625 7 42 2.174(-9) 28 48.2 6.754(-9)
Table 4: Results from Algorithm 4.1 on function f1. NC is the number of coarse grid points
and σ is the average number of non-zero elements per column in the approximation to B.
[2] R. K. Beatson, J. B. Cherrie and C. T. Mouat, Fast fitting of radial basis func-
tions: Methods based on preconditioned GMRES iteration, Advances in Computa-
tional Mathematics, 11 (1999), 253-270.
[3] R. K. Beatson, J. B. Cherrie and G. N. Newsam, Fast evaluation of radial basis
functions: Methods for generalised multiquadrics in Rn. Manuscript (2001).
[4] R. K. Beatson and C. T. Mouat, Fast Kriging, University of Canterbury Research
Report number UCDMS2000/2, (2000).
20 rbf collocation
[5] R. K. Beatson and G. N. Newsam, Fast evaluation of radial basis functions: I, Com-
puters and Mathematics with Applications, 24, No. 12 (1992), 7-19.
[6] R. E. Carlson and T. A. Foley, The parameter R2 in multiquadric interpolation, Com-
puters and Mathematics with Applications, 21 (1991), 29-42.
[7] G. E. Fasshauer, On the numerical solution of differential equations with radial basis
functions. Boundary Element Technology XIII, C. S. Chen, C. A. Brebbia, and D. W.
Pepper (eds), WIT Press, (1999), 291-300.
[8] G. E. Fasshauer, Solving partial differential equations by collocation with radial basis
functions. Surface Fitting and Multiresolution Methods, A. Le Mehaute, C. Rabut,
and L. L. Schumaker (eds.), Vanderbilt University Press, (1997), 131-138.
[9] G. E. Fasshauer and J. W. Jerome, Multistep approximation algorithms: improved
convergence rates through postconditioning with smoothing kernels.Advances in Com-
putational Mathematics, 10 (1999), 1-27.
[10] G. E. Fasshauer, Nonsymmetric multilevel RBF collocation within an operator newton
framework for nonlinear PDEs, to appear in Trends in Approximation Theory.
[11] G. E. Fasshauer, E. C. Gartland and J. W. Jerome, Newton iteration for partial
differential equations and the approximation of the identity, Numerical Algorithms,
25 (2000), 181-195.
[12] R. Franke, Scattered data interpolation: tests of some methods, Mathematics of Com-
putation, 38 (1982), 181-200.
[13] C. Franke and R. Schaback, Convergence order estimates of meshless collocation meth-
ods using radial basis functions, Advances in Computational Mathematics, 8 (1998),
381-399.
[14] C. Franke and R. Schaback, Solving partial differential equations by collocation using
radial basis functions, Applied Mathematics and Computation, 93 (1998), 73-82.
c. t. mouat and r. k. beatson 21
[15] Y. C. Hon, A RBFs method for solving options pricing model, Proceedings of Advances
in Scientific Computing and Modeling, Alicante, Spain (1998), 210-230.
[16] Y. C. Hon and Z. Wu, Error estimation on using radial basis functions for solving
option pricing models. Manuscript.
[17] E. J. Kansa, Multiquadrics - a scattered data approximation scheme with applications
to computational fluid-dynamics-II, Computers and Mathematics with Applications,
19 (1990), 147-161.
[18] E. J. Kansa and Y. C. Hon, Circumventing the ill-conditioining problem with mul-
tiquadric radial basis functions: applications to elliptic partial differential equations,
Computers and Mathematics with Applications, 39 (2000), 123-137.
[19] W. R. Madych, Miscellaneous error bounds for multiquadric and related interpolators,
Computers and Mathematics with Applications, 24 (1992), 121-138.
[20] R. Schaback and Y. C. Hon, On unsymmetric collocation by radial basis functions,
Journal of Applied Mathematics and Computation, 119 (2001), 177-186.
[21] H. Wendland, Piecewise polynomial, positive definite and compactly supported radial
functions of minimal degree, Advances in Computational Mathematics, 4 (1995), 389-
396.
[22] Z. Wu, Hermite-Birkhoff interpolation of scattered data by radial basis functions,
Approximation Theory and its Applications, 8 (1992), 1-10.